/ROB535-LiDAR_only_objectdetection

ROB535 final project Team 16

Primary LanguagePython

ROB535-LiDAR_only_objectdetection

Our ROB535 final project focus on comparing performance of three different 3D detection models on nuScenes dataset. The voxelnext is implemented based on OpenPCDet, the PointPillars and CenterPoint are implemented based on the mmdetection3D. All models evaluated on the validation set of nuScenes V1.0.

  1. Installation Follow the instruction in VoxelNext and mmdetection3D to install required environment

  2. Data Preparation Download nuScenes dataset from official website.

    Install the nuscenes-devkit with version 1.0.5 by running the following command:

    pip install nuscenes-devkit==1.0.5

    For OpenPCDet

    python -m pcdet.datasets.nuscenes.nuscenes_dataset --func create_nuscenes_infos \
 --cfg_file tools/cfgs/dataset_configs/nuscenes_dataset.yaml \
 --version v1.0-trainval

    For mmdetection3D

    python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

  3. Download pretrained weight from official repo

  4. Run evaluate python file For VoxelNext

    bash VoxelNeXt/tools/scripts/dist_test.sh 4 --cfg_file VoxelNeXt/tools/cfgs/nuscenes_models/cbgs_voxel0075_voxelnext.yaml --ckpt path/to/pretain_model

    For CenterPoint and PointPillar

    CONFIG_FILE=mmdetection3d/configs/centerpoint/centerpoint_0075voxel_second_secfpn_dcn_4x8_cyclic_20e_nus.py
CONFIG_FILE=mmdetection3d/configs/pointpillars/hv_pointpillars_secfpn_sbn-all_4x8_2x_nus-3d.py
/tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} --format-only --eval-options 'jsonfile_prefix=${OUTPUT_PREFIX}'

    evalution metrics in nuscenes format could be gained by run

    python eval_metric.py